The Research On Fast Detection Algorithm For Small Targets Of Aerial Images Based On SSD

Aerial target detection has become one of the hot topics in Computer Vision.Object detection in aerial images has broad application prospects in accurate military target detection and fire detection.Due to small targets and complex backgrounds,aerial target detection has always been a challenge.This paper builds a deep learning platform,proposes an aerial target detection method based on the SSD(Single Shot multibox Detector)algorithm,and improves the SSD algorithm to achieve fast small target detection in aerial images.The main work of this paper is in the following areas:(1)Build a deep learning platform.This paper uses Caffe deep learning framework to study aerial target detection.This paper builds the Caffe platform used by many people in the LAN,and writes scripts for data analysis such as network structure analysis and feature map display.(2)Chose SSD algorithm as the aerial target detection algorithm,and proposed an improved SSD algorithm based on feature fusion,named F-SSD.By analyzing the difficulty of target detection in aerial image,this paper proposes the aerial target detection based on the SSD for achieving both effective and fast target detection.After analyzing the SSD network,this paper finds that low-level features include more small-target information and less semantic information,and higher-level features include more semantic information and fewer smalltarget information.F-SSD combines the low-level features with high-level features to improve SSD's the performance on small object detection.F-SSD achieves 93.7% mAP by training and testing the actual aerial image and for 300?300 input,which is 4.05% higher than SSD.(3)In order to enhance the SSD target detection capability,this paper proposes an improved SSD-based algorithm based on feature fusion and enhancement,named EF-SSD.The feature contains many channels,and the characteristics of different channels contribute differently to the target detection result.The suppression of invalid channel features based on the enhancement of effective channel features can effectively improve the performance of target detection.Based on F-SSD,EF-SSD uses feature enhancement methods to improve F-SSD.In terms of detection accuracy,by training and testing the actual aerial image of the resolution,the mAP of the EF-SSD reached 94.92%,which is 1.22% higher than that of the F-SSD,and 5.27% higher than that of the SSD.On the VOC2007 test set,the EF-SSD reached 79.3% mAP,higher than the mAP of the F-SSD and the SSD.In terms of speed,the EF-SSD with 300?300 input detectes an image on the K80 graphics card for 51.2 ms,an increase of 8.9% compared to the SSD.This paper builds a deep learning platform for the study of aerial image target detection and chooses SSD algorithm as the aerial target detection algorithm.In order to improve the detection effect of small targets and the overall target detection effect,this paper proposes an improved SSD-based algorithm based on feature fusion and enhancement.The experimental results show that the improved SSD-based algorithm can fast achieve small target detection of aerial images,which has a very good application value.